{"id":542,"date":"2025-02-10T11:00:19","date_gmt":"2025-02-10T11:00:19","guid":{"rendered":"https:\/\/pinglex.ai\/?p=542"},"modified":"2025-02-10T03:18:47","modified_gmt":"2025-02-10T03:18:47","slug":"%e3%80%8a%e5%a4%9a%e9%9a%8e%e6%ae%b5%e5%bc%b7%e5%8c%96%e5%ad%b8%e7%bf%92%e8%a8%93%e7%b7%b4%e5%a6%82%e4%bd%95%e6%8f%90%e5%8d%87-deepseek-r1-zero-%e7%9a%84%e8%87%aa%e6%88%91%e4%bf%ae%e6%ad%a3%e8%83%bd","status":"publish","type":"post","link":"https:\/\/pinglex.ai\/?p=542","title":{"rendered":"\u300a\u591a\u968e\u6bb5\u5f37\u5316\u5b78\u7fd2\u8a13\u7df4\u5982\u4f55\u63d0\u5347 DeepSeek-R1-Zero \u7684\u81ea\u6211\u4fee\u6b63\u80fd\u529b\uff1f\u300b"},"content":{"rendered":"\n<p class=\"\">\uff08\u8207Felo Research\u591a\u8f2a\u5c0d\u8a71\u5f8c\u751f\u6210\u7684\u63d0\u793a\u8a5e\uff09<\/p>\n\n\n\n<p class=\"\">\u591a\u968e\u6bb5\u5f37\u5316\u5b78\u7fd2\u8a13\u7df4\u662f DeepSeek-R1-Zero \u5c55\u73fe\u81ea\u6211\u4fee\u6b63\u80fd\u529b\u7684\u6838\u5fc3\u6280\u8853\u4e4b\u4e00\u3002\u9019\u7a2e\u8a13\u7df4\u65b9\u6cd5\u900f\u904e\u5206\u968e\u6bb5\u7684\u8a2d\u8a08\uff0c\u9010\u6b65\u63d0\u5347\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3001\u81ea\u6211\u6aa2\u67e5\u80fd\u529b\u4ee5\u53ca\u932f\u8aa4\u4fee\u6b63\u80fd\u529b\u3002\u4ee5\u4e0b\u662f\u591a\u968e\u6bb5\u5f37\u5316\u5b78\u7fd2\u5982\u4f55\u5be6\u73fe\u9019\u4e00\u76ee\u6a19\u7684\u8a73\u7d30\u89e3\u6790\uff1a<\/p>\n\n\n\n<p class=\"\">1. \u591a\u968e\u6bb5\u5f37\u5316\u5b78\u7fd2\u7684\u6838\u5fc3\u8a2d\u8a08<\/p>\n\n\n\n<p class=\"\"><a><\/a>DeepSeek-R1-Zero \u7684\u591a\u968e\u6bb5\u5f37\u5316\u5b78\u7fd2\u8a13\u7df4\u6d41\u7a0b\u65e8\u5728\u8b93\u6a21\u578b\u9010\u6b65\u638c\u63e1\u63a8\u7406\u8207\u81ea\u6211\u4fee\u6b63\u80fd\u529b\uff0c\u4e26\u89e3\u6c7a\u50b3\u7d71\u8a13\u7df4\u65b9\u6cd5\u7684\u4e0d\u8db3\u3002\u5176\u4e3b\u8981\u7279\u9ede\u5305\u62ec\uff1a<\/p>\n\n\n\n<p class=\"\">\u7d14\u5f37\u5316\u5b78\u7fd2\u9a45\u52d5\uff1aDeepSeek-R1-Zero \u5b8c\u5168\u4f9d\u8cf4\u5f37\u5316\u5b78\u7fd2\uff08Reinforcement Learning, RL\uff09\uff0c\u4e0d\u4f7f\u7528\u76e3\u7763\u5fae\u8abf\uff08Supervised Fine-Tuning, SFT\uff09\uff0c\u76f4\u63a5\u5f9e\u9810\u8a13\u7df4\u6a21\u578b\uff08DeepSeek-V3-Base\uff09\u9032\u884c\u8a13\u7df4\u3002<\/p>\n\n\n\n<p class=\"\">\u7fa4\u7d44\u76f8\u5c0d\u7b56\u7565\u512a\u5316\uff08Group Relative Policy Optimization, GRPO\uff09\uff1a\u9019\u662f\u4e00\u7a2e\u5c08\u70ba DeepSeek \u8a2d\u8a08\u7684\u5f37\u5316\u5b78\u7fd2\u7b97\u6cd5\uff0c\u901a\u904e\u5c0d\u6bd4\u540c\u4e00\u554f\u984c\u7684\u591a\u500b\u89e3\u7b54\uff0c\u6839\u64da\u9810\u5b9a\u898f\u5247\uff08\u5982\u6e96\u78ba\u6027\u548c\u683c\u5f0f\uff09\u5206\u914d\u734e\u52f5\uff0c\u5f9e\u800c\u5f15\u5c0e\u6a21\u578b\u751f\u6210\u66f4\u9ad8\u8cea\u91cf\u7684\u8f38\u51fa\u3002<\/p>\n\n\n\n<p class=\"\">2. \u591a\u968e\u6bb5\u8a13\u7df4\u5982\u4f55\u63d0\u5347\u81ea\u6211\u4fee\u6b63\u80fd\u529b<\/p>\n\n\n\n<p class=\"\">\u591a\u968e\u6bb5\u5f37\u5316\u5b78\u7fd2\u7684\u8a2d\u8a08\u8b93\u6a21\u578b\u80fd\u5920\u9010\u6b65\u5b78\u6703\u6aa2\u67e5\u548c\u4fee\u6b63\u81ea\u5df1\u7684\u8f38\u51fa\uff0c\u4ee5\u4e0b\u662f\u5177\u9ad4\u7684\u5be6\u73fe\u65b9\u5f0f\uff1a<\/p>\n\n\n\n<p class=\"\">(1) \u734e\u52f5\u4fe1\u865f\u9a45\u52d5\u7684\u81ea\u6211\u4fee\u6b63<\/p>\n\n\n\n<p class=\"\">\u6e96\u78ba\u6027\u734e\u52f5\uff1a\u6a21\u578b\u5728\u751f\u6210\u7b54\u6848\u6642\uff0c\u6703\u6839\u64da\u5176\u6b63\u78ba\u6027\u7372\u5f97\u734e\u52f5\u3002\u4f8b\u5982\uff0c\u5728\u6578\u5b78\u63a8\u7406\u4e2d\uff0c\u6a21\u578b\u9700\u8981\u751f\u6210\u7b26\u5408\u898f\u5247\u7684\u6b63\u78ba\u7b54\u6848\uff0c\u5426\u5247\u6703\u53d7\u5230\u61f2\u7f70\u3002\u9019\u7a2e\u6a5f\u5236\u4fc3\u4f7f\u6a21\u578b\u5728\u751f\u6210\u904e\u7a0b\u4e2d\u4e3b\u52d5\u6aa2\u67e5\u6bcf\u4e00\u6b65\u63a8\u7406\u662f\u5426\u6b63\u78ba\u3002<\/p>\n\n\n\n<p class=\"\">\u683c\u5f0f\u734e\u52f5\uff1a\u6a21\u578b\u9700\u4ee5\u6e05\u6670\u7684\u683c\u5f0f\u8868\u9054\u5176\u63a8\u7406\u904e\u7a0b\uff0c\u4f8b\u5982\u4f7f\u7528 &lt;think&gt; \u548c &lt;\/think&gt; \u6a19\u7c64\u4f86\u6a19\u8a18\u601d\u7dad\u93c8\u3002\u9019\u4e0d\u50c5\u63d0\u5347\u4e86\u8f38\u51fa\u7684\u53ef\u8b80\u6027\uff0c\u4e5f\u8b93\u6a21\u578b\u66f4\u5bb9\u6613\u6aa2\u67e5\u81ea\u5df1\u7684\u908f\u8f2f\u662f\u5426\u4e00\u81f4\u3002<\/p>\n\n\n\n<p class=\"\">(2) \u81ea\u6211\u53cd\u601d\u8207\u300c\u9813\u609f\u6642\u523b\u300d<\/p>\n\n\n\n<p class=\"\">\u81ea\u6211\u9a57\u8b49\u6a5f\u5236\uff1a\u6a21\u578b\u5728\u63a8\u7406\u904e\u7a0b\u4e2d\u6703\u4e3b\u52d5\u6aa2\u67e5\u5df2\u751f\u6210\u7684\u5167\u5bb9\u662f\u5426\u7b26\u5408\u908f\u8f2f\u3002\u4f8b\u5982\uff0c\u5728\u89e3\u6c7a\u908f\u8f2f\u8b0e\u984c\u6642\uff0c\u6a21\u578b\u6703\u56de\u6eaf\u4e26\u91cd\u65b0\u8a55\u4f30\u5176\u63a8\u7406\u6b65\u9a5f\uff0c\u78ba\u4fdd\u6700\u7d42\u7b54\u6848\u6b63\u78ba\u3002<\/p>\n\n\n\n<p class=\"\">\u300c\u9813\u609f\u6642\u523b\u300d\uff08Aha Moment\uff09\uff1a\u5728\u5f37\u5316\u5b78\u7fd2\u904e\u7a0b\u4e2d\uff0c\u6a21\u578b\u81ea\u7136\u6e67\u73fe\u51fa\u4e00\u7a2e\u81ea\u6211\u53cd\u601d\u7684\u80fd\u529b\uff0c\u80fd\u5920\u5728\u767c\u73fe\u932f\u8aa4\u5f8c\u91cd\u65b0\u601d\u8003\u4e26\u4fee\u6b63\u3002\u4f8b\u5982\uff0c\u5728\u89e3\u6c7a\u300c\u8fb2\u592b\u904e\u6cb3\u300d\u554f\u984c\u6642\uff0c\u6a21\u578b\u6700\u521d\u751f\u6210\u4e86\u4e00\u500b\u932f\u8aa4\u7684\u89e3\u6cd5\uff0c\u4f46\u96a8\u5f8c\u901a\u904e\u91cd\u65b0\u8a55\u4f30\u908f\u8f2f\uff0c\u627e\u5230\u4e86\u6b63\u78ba\u7684\u89e3\u6cd5\u3002<\/p>\n\n\n\n<p class=\"\">(3) \u52d5\u614b\u8cc7\u6e90\u5206\u914d<\/p>\n\n\n\n<p class=\"\">\u5206\u914d\u66f4\u591a\u8a08\u7b97\u8cc7\u6e90\uff1a\u6a21\u578b\u5728\u8655\u7406\u8907\u96dc\u554f\u984c\u6642\uff0c\u6703\u81ea\u52d5\u5206\u914d\u66f4\u591a\u7684\u300c\u601d\u8003\u6642\u9593\u300d\u4f86\u9032\u884c\u6df1\u5ea6\u63a8\u7406\uff0c\u9019\u7a2e\u884c\u70ba\u662f\u5f37\u5316\u5b78\u7fd2\u904e\u7a0b\u4e2d\u81ea\u7136\u6e67\u73fe\u7684\u7d50\u679c\u3002<\/p>\n\n\n\n<p class=\"\">3. \u591a\u968e\u6bb5\u8a13\u7df4\u7684\u5177\u9ad4\u6d41\u7a0b<\/p>\n\n\n\n<p class=\"\">DeepSeek-R1-Zero \u7684\u591a\u968e\u6bb5\u5f37\u5316\u5b78\u7fd2\u6d41\u7a0b\u5982\u4e0b\uff1a<\/p>\n\n\n\n<p class=\"\">\u521d\u59cb\u7b56\u7565\u751f\u6210\uff1a\u57fa\u65bc DeepSeek-V3-Base \u6a21\u578b\uff0c\u4f7f\u7528 GRPO \u7b97\u6cd5\u751f\u6210\u521d\u59cb\u7b56\u7565\uff0c\u4e26\u901a\u904e\u591a\u6b21\u8a66\u932f\u5b78\u7fd2\u5982\u4f55\u89e3\u6c7a\u554f\u984c\u3002<\/p>\n\n\n\n<p class=\"\">\u591a\u6a23\u5316\u8f38\u51fa\u8207\u5c0d\u6bd4\u5b78\u7fd2\uff1a\u5c0d\u540c\u4e00\u554f\u984c\u751f\u6210\u591a\u500b\u89e3\u7b54\uff0c\u4e26\u6839\u64da\u9810\u5b9a\u898f\u5247\uff08\u5982\u6e96\u78ba\u6027\u548c\u683c\u5f0f\uff09\u5206\u914d\u734e\u52f5\uff0c\u6a21\u578b\u5b78\u7fd2\u5982\u4f55\u751f\u6210\u66f4\u9ad8\u8cea\u91cf\u7684\u7b54\u6848\u3002<\/p>\n\n\n\n<p class=\"\">\u53cd\u8986\u512a\u5316\uff1a\u6a21\u578b\u5728\u6bcf\u6b21\u751f\u6210\u7b54\u6848\u5f8c\uff0c\u6703\u6839\u64da\u734e\u52f5\u4fe1\u865f\u8abf\u6574\u5176\u7b56\u7565\uff0c\u9010\u6b65\u63d0\u5347\u63a8\u7406\u80fd\u529b\u548c\u81ea\u6211\u4fee\u6b63\u80fd\u529b\u3002<\/p>\n\n\n\n<p class=\"\">4. \u5be6\u969b\u61c9\u7528\u4e2d\u7684\u8868\u73fe<\/p>\n\n\n\n<p class=\"\">\u6578\u5b78\u63a8\u7406\uff1a\u5728\u6578\u5b78\u554f\u984c\u4e2d\uff0cDeepSeek-R1-Zero \u80fd\u5920\u6aa2\u67e5\u6bcf\u4e00\u6b65\u8a08\u7b97\uff0c\u4e26\u5728\u767c\u73fe\u932f\u8aa4\u6642\u91cd\u65b0\u8a08\u7b97\u3002\u4f8b\u5982\uff0c\u5728 AIME \u6e2c\u8a66\u4e2d\uff0c\u6a21\u578b\u901a\u904e\u591a\u6b21\u81ea\u6211\u4fee\u6b63\u5c07\u6b63\u78ba\u7387\u986f\u8457\u63d0\u5347\u3002<\/p>\n\n\n\n<p class=\"\">\u908f\u8f2f\u63a8\u7406\uff1a\u6a21\u578b\u80fd\u5920\u6aa2\u67e5\u5176\u908f\u8f2f\u63a8\u7406\u662f\u5426\u4e00\u81f4\uff0c\u4e26\u5728\u5fc5\u8981\u6642\u91cd\u65b0\u751f\u6210\u90e8\u5206\u5167\u5bb9\u3002\u4f8b\u5982\uff0c\u5728\u89e3\u6c7a\u8907\u96dc\u908f\u8f2f\u554f\u984c\u6642\uff0c\u6a21\u578b\u6703\u5148\u751f\u6210\u521d\u6b65\u7b54\u6848\uff0c\u7136\u5f8c\u6aa2\u67e5\u5176\u5408\u7406\u6027\uff0c\u82e5\u767c\u73fe\u4e0d\u4e00\u81f4\u5247\u9032\u884c\u4fee\u6b63\u3002<\/p>\n\n\n\n<p class=\"\">5. \u6311\u6230\u8207\u672a\u4f86\u6539\u9032\u65b9\u5411<\/p>\n\n\n\n<p class=\"\">\u5118\u7ba1\u591a\u968e\u6bb5\u5f37\u5316\u5b78\u7fd2\u986f\u8457\u63d0\u5347\u4e86 DeepSeek-R1-Zero \u7684\u81ea\u6211\u4fee\u6b63\u80fd\u529b\uff0c\u4f46\u4ecd\u5b58\u5728\u4e00\u4e9b\u6311\u6230\uff1a<\/p>\n\n\n\n<p class=\"\">\u8f38\u51fa\u53ef\u8b80\u6027\u554f\u984c\uff1a\u6a21\u578b\u751f\u6210\u7684\u63a8\u7406\u904e\u7a0b\u53ef\u80fd\u904e\u65bc\u5197\u9577\u6216\u6df7\u4e82\uff0c\u5f71\u97ff\u7528\u6236\u9ad4\u9a57\u3002<\/p>\n\n\n\n<p class=\"\">\u8a9e\u8a00\u6df7\u96dc\u554f\u984c\uff1a\u5728\u591a\u8a9e\u8a00\u74b0\u5883\u4e2d\uff0c\u6a21\u578b\u53ef\u80fd\u6703\u6df7\u7528\u4e0d\u540c\u8a9e\u8a00\uff0c\u5c0e\u81f4\u8868\u9054\u4e0d\u4e00\u81f4\u3002<\/p>\n\n\n\n<p class=\"\">\u70ba\u4e86\u89e3\u6c7a\u9019\u4e9b\u554f\u984c\uff0c\u5f8c\u7e8c\u7248\u672c\uff08\u5982 DeepSeek-R1\uff09\u5f15\u5165\u4e86\u51b7\u555f\u52d5\u6578\u64da\u548c\u591a\u968e\u6bb5\u8a13\u7df4\u6d41\u7a0b\uff0c\u9032\u4e00\u6b65\u63d0\u5347\u4e86\u6a21\u578b\u7684\u63a8\u7406\u6027\u80fd\u8207\u8868\u9054\u80fd\u529b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" width=\"1024\" height=\"408\" data-src=\"https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-10-1024x408.png\" alt=\"\" class=\"wp-image-543 lazyload\" data-srcset=\"https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-10-1024x408.png 1024w, https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-10-300x120.png 300w, https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-10-768x306.png 768w, https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-10-600x239.png 600w, https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-10.png 1196w\" data-sizes=\"(max-width: 1024px) 100vw, 1024px\" src=\"data:image\/svg+xml;base64,PHN2ZyB3aWR0aD0iMSIgaGVpZ2h0PSIxIiB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciPjwvc3ZnPg==\" style=\"--smush-placeholder-width: 1024px; --smush-placeholder-aspect-ratio: 1024\/408;\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" width=\"1024\" height=\"365\" data-src=\"https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-11-1024x365.png\" alt=\"\" class=\"wp-image-544 lazyload\" data-srcset=\"https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-11-1024x365.png 1024w, https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-11-300x107.png 300w, https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-11-768x274.png 768w, https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-11-600x214.png 600w, https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-11.png 1172w\" data-sizes=\"(max-width: 1024px) 100vw, 1024px\" src=\"data:image\/svg+xml;base64,PHN2ZyB3aWR0aD0iMSIgaGVpZ2h0PSIxIiB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciPjwvc3ZnPg==\" style=\"--smush-placeholder-width: 1024px; --smush-placeholder-aspect-ratio: 1024\/365;\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" width=\"1024\" height=\"374\" data-src=\"https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-12-1024x374.png\" alt=\"\" class=\"wp-image-545 lazyload\" data-srcset=\"https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-12-1024x374.png 1024w, https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-12-300x110.png 300w, https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-12-768x281.png 768w, https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-12-1536x561.png 1536w, https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-12-600x219.png 600w, https:\/\/pinglex.ai\/wp-content\/uploads\/2025\/02\/image-12.png 1752w\" data-sizes=\"(max-width: 1024px) 100vw, 1024px\" src=\"data:image\/svg+xml;base64,PHN2ZyB3aWR0aD0iMSIgaGVpZ2h0PSIxIiB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciPjwvc3ZnPg==\" style=\"--smush-placeholder-width: 1024px; --smush-placeholder-aspect-ratio: 1024\/374;\" \/><\/figure>\n","protected":false},"excerpt":{"rendered":"<p>\uff08\u8207Felo Research\u591a\u8f2a\u5c0d\u8a71\u5f8c\u751f\u6210\u7684\u63d0\u793a\u8a5e\uff09 \u591a\u968e\u6bb5\u5f37\u5316\u5b78\u7fd2\u8a13\u7df4\u662f DeepSeek-R1-Zero [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":544,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"nf_dc_page":"","iawp_total_views":1,"footnotes":""},"categories":[56,27],"tags":[],"class_list":["post-542","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-56","category-llm"],"_links":{"self":[{"href":"https:\/\/pinglex.ai\/index.php?rest_route=\/wp\/v2\/posts\/542","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/pinglex.ai\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/pinglex.ai\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/pinglex.ai\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/pinglex.ai\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=542"}],"version-history":[{"count":1,"href":"https:\/\/pinglex.ai\/index.php?rest_route=\/wp\/v2\/posts\/542\/revisions"}],"predecessor-version":[{"id":546,"href":"https:\/\/pinglex.ai\/index.php?rest_route=\/wp\/v2\/posts\/542\/revisions\/546"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/pinglex.ai\/index.php?rest_route=\/wp\/v2\/media\/544"}],"wp:attachment":[{"href":"https:\/\/pinglex.ai\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=542"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/pinglex.ai\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=542"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/pinglex.ai\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=542"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}